Download Report

Image Segmentation in Twenty Questions
Christian Rupprecht1,2
1
Lo¨ıc Peter1
Nassir Navab1,2
Technische Universit¨at M¨unchen, Munich, Germany
2
Johns Hopkins University, Baltimore MD, USA
{christian.rupprecht, peter, navab}@in.tum.de
Abstract
Consider the following scenario between a human user
and the computer. Given an image, the user thinks of an
object to be segmented within this picture, but is only allowed to provide binary inputs to the computer (yes or no).
In these conditions, can the computer guess this hidden segmentation by asking well-chosen questions to the user? We
introduce a strategy for the computer to increase the accuracy of its guess in a minimal number of questions. At
each turn, the current belief about the answer is encoded
in a Bayesian fashion via a probability distribution over
the set of all possible segmentations. To efficiently handle
this huge space, the distribution is approximated by sampling representative segmentations using an adapted version of the Metropolis-Hastings algorithm, whose proposal
moves build on a geodesic distance transform segmentation
method. Following a dichotomic search, the question halving the weighted set of samples is finally picked, and the
provided answer is used to update the belief for the upcoming rounds. The performance of this strategy is assessed on
three publicly available datasets with diverse visual properties. Our approach shows to be a tractable and very adaptive solution to this problem.
1. Introduction
Twenty Questions is a classical two-player game involving a questioner and an oracle. Before the game starts, the
oracle thinks about something (e.g. an object, an animal or a
character) which we will call the answer, and the questioner
is allowed to ask a series of binary questions to guess what
the answer is. While the game originally involves a human in the role of the oracle and another human as player,
the development of artificial intelligence techniques led to
softwares where the computer tries to guess what the human user has in mind [1, 2]. Going in the same direction,
we consider in this work the case of a computer playing the
Twenty Questions game with a human user in the role of the
oracle, where the expected answer is a binary segmentation
of a given image. We introduce a strategy for the computer
to provide a guess as accurate as possible in a limited number of questions. Alternatively, the proposed approach can
also be seen as an interactive segmentation task. The crucial characteristic of our scenario is that the user interaction
with the machine is restricted to binary inputs (yes/no) instead of the usual scribbles or bounding boxes [10]. This
setting is common to provide relevant feedback for interactive image retrieval [19, 24, 26] but was never considered
in the context of segmentation. Moreover, if combined with
e.g. a voice recognition system, our method can eventually
provide a hands-free segmentation technique, for which an
alternative solution was proposed by Sadeghi et al. [21] who
used an eye tracker for seed placement. This has potential
applications like segmentation of medical data in sterilized
operating room environments, where physical interactions
with objects have to be avoided.
Our method can be summarized as follows. At each turn,
the question to be posed consists in asking whether a wellchosen pixel is inside the hidden segmentation or not. The
choice of the location to ask about is made as to maximize
the information brought by the answer. More precisely, we
define for any possible segmentation a probability to have
been picked by the oracle. Each segmentation score is based
on two aspects: its intrinsic quality in terms of image partitioning (i.e. the homogeneity of the segmented object),
and its compatibility with the answers already collected.
Hence, this defines a probability distribution over the set
of possible segmentations S. In theory, the ideal question
choice would follow a divide and conquer approach and
halve S in two parts of approximately equal probabilities.
However, the huge size of S excludes the exhaustive computation of the aforementioned distribution, which we approximate instead by sampling a series of segmentations
proportionally to their probability of having been chosen
by the oracle. This sampling is performed via a Markov
Chain Monte Carlo (MCMC) framework, which builds on
the Metropolis-Hastings algorithm [18] and consists in ex-
Figure 1: Overview of the Twenty Questions segmentation scenario. Given an image, the object to segment is secretly
chosen by the human user. At every step, the computer asks whether a certain pixel is located inside the desired segmentation.
After a predefined number of questions, it returns its guess about the answer.
ploring S while spending more time in areas of high probability. The most informative question can be computed
from the drawn samples in a tractable way, and the knowledge brought by the answer is included in the probabilistic
model for the future rounds. We evaluate our method on
three publicly available segmentation datasets with various
visual properties and compare it with several baselines. The
experiments demonstrate that our approach is a promising
solution to the problem of interactive object segmentation
with binary inputs only. The question selection strategy is
fast enough (around 1 second) to be incorporated in a software in practice and does not rely on any offline training
step, which also makes the framework overall very adaptive
to different types of images.
keep the human user in the loop and suggest the most informative areas to label next, in an active learning fashion.
An important aspect of our approach is the intrinsic ambiguity of the image parsing task, as one cannot anticipate
the semantic level of the segmentation picked by the oracle. In this direction, Tu and Zhu [27] introduced a datadriven MCMC framework based on active contours able
to generate several parsings of a same image. Recent alternatives identify a set of candidate relevant objects in a
scene [6, 7, 12, 15] by learning plausible object appearances or shapes. In our case, no offline training is performed
so that the method can be applied to any kind of image or
ground truth.
2. Methods
Related Work In this paragraph, we briefly review the
connections of our work with existing methods. The Twenty
Questions setting has already been mentioned in the computer vision community through the work of Branson et
al. [5] and its extension by Wah et al. [28] for interactive
fine-grained image classification. These works consider the
case where both the human user and the machine ignore
the image label, but (i) the machine knows which questions
are important to ask to find out the answer and (ii) the human is able to answer these questions which are based on
the visual aspect of the scene. Hence, combining the expertise of the computer with the visual abilities of the human
allows to find collaboratively the hidden image label. Beyond the differences in terms of task (image classification vs
segmentation), this setting is fundamentally different from
ours, where the object to be segmented is perfectly known
by the human user and has to be guessed by the computer.
Interactive segmentation techniques usually rely on
seeds [4, 10, 20] or bounding boxes [11, 16] that are manually placed by a human user in or around the object of interest. Closer to our work, a few approaches [3, 8, 13, 14, 23]
2.1. Problem Statement and Notations
An image I is defined over a lattice Ω = {1, . . . , h} ×
{1, . . . , w} where h and w respectively denote the height
and width of I. We define a binary segmentation of the
image I as a function s : Ω → {0, 1} indicating whether
the pixel (x, y) of the image is inside (s(x, y) = 1) or outside (s(x, y) = 0) the segmented area. Depending on the
context, s can also be seen as a vector s ∈ S = {0, 1}|Ω| .
Our problem can be formalized as follows. Initially,
given a fixed image I, the oracle (i.e. the human user) decides on a segmentation ˆs ∈ S that has to be found by the
computer. To do so, the computer is going to ask to the oracle a series of binary questions of the form Q(p): “Is the displayed location p inside the object you want to segment?”.
Choosing the best question to ask amounts to finding the
most informative location p. In return, the answer directly
provides the true label l(p) ∈ {0, 1} at this location.
After k questions have been posed, the collected answers
provide two reliable sets Σk− and Σ k+ of
and
background
foreground seeds respectively, with Σk− + Σk+ = k. We
also denote Σk = Σk− ∪ Σk+ the set of reliable seeds collected. This knowledge is encoded through a Bayesian posterior probability p(s = ˆs|Σk ) over the set of segmentations
S stating how likely it is that the segmentation s has been
initially picked by the user given the known seeds revealed
by the answers already collected. We will denote this probability p(s|Σk ) in the rest of the paper to make clear that
this probability is seen as a function of s and as a probability distribution over S. If this posterior could be computed
for every possible segmentation in S, an optimal divide and
conquer strategy would halve at each turn the set of possible segmentations into two subsets of probability 0.5 each.
However, the set S has an extremely large size, with theoretically 2|Ω| possibilities, which excludes the exhaustive
computation of p(s|Σk ) over the whole set S. To overcome this, we propose at each iteration k to approximate
the posterior p(s|Σk ) by a series of samples sk1 , . . . , skN ∈ S
drawn according to a Markov Chain Monte Carlo (MCMC)
scheme, which is described in detail in Sec. 2.2. After these
N samples have been drawn, we select the most informative
question based on these samples by following the question
selection method exposed in Sec. 2.3. These two sampling
and question selection steps are then iterated until a predefined amount of allowed questions is reached.
Input : Image I, number of allowed questions K
Σ0− ← ∅; Σ0+ ← ∅;
for k ← 0 to K − 1 do
Sample sk1 , . . . , skN from p(.|Σk );
Find most informative location pk w.r.t sk1 , . . . , skN ;
Ask question Q(pk ) and receive true label l(pk );
if l(pk ) = 0 then
Σk+1
← Σk− ∪ {pk };
−
k+1
Σ+ ← Σk+ ;
else
Σk+1
← Σk− ;
−
k+1
Σ+ ← Σk+ ∪ {pk };
end
end
G ← GDT-segmentation with seeds ΣK ;
Output: Guess G of the oracle segmentation
Algorithm 1: Overview of our method. At each step,
the question selection strategy consists in sampling segmentations according to their probability of having been
picked by the oracle, and asking the question related to
the most informative location with respect to these samples. This location together with the provided label form a
seed which is added to either Σ0− or Σ0+ depending on its
label.
2.2. Sampling Likely Segmentations with MCMC
In this section, we introduce our procedure for sampling
representative segmentations from the posterior probability
distribution p(.|Σk ) conditioned on the current knowledge.
Following the classical Metropolis-Hastings algorithm [18]
and an original idea from Tu and Zhu [27] introduced in the
context of image parsing into multiple regions, we define
our sampling procedure as a Markov chain over the space
of segmentations with transition probabilities defined as follows. Given a current segmentation s, a new segmentation
candidate s0 is suggested according to a proposal distribution q(s0 |Σk , s) to be the new state of the Markov chain.
This move is accepted with probability min(1, α) with
α=
p(s0 |Σk )q(s|Σk , s0 )
,
p(s|Σk )q(s0 |Σk , s)
(1)
where p(.|Σk ) denotes the posterior probability distribution
according to which we want to draw samples, i.e. the probability for a segmentation to have been picked by the user
given the currently known seeds Σk . Starting from an initial
segmentation sk0 , a succession of segmentations is generated
and the ski , 1 ≤ i ≤ N are selected as samples at a fixed
rate during this exploration process. Additionally, a burn-in
step is performed before starting collecting samples so that
the dependency on the initial state sk0 is strongly reduced.
To complete the description of the segmentation sampling
method, one has to carefully define the posterior distribution and the proposal distribution. Our design for these distributions will be described in Sec. 2.2.2 and Sec. 2.2.3 respectively. They both build on a parametrization of the set
of segmentations S which we are now going to introduce.
2.2.1
State-Space Parametrization
To facilitate the design of efficient proposal distributions in
Eq. 1, we need a way to deform a segmentation into another.
In their data-driven MCMC framework for image parsing,
Tu and Zhu proposed to deform active contours [27]. In our
case, the MCMC paradigm takes place between two questions in the context of a human/machine interaction. Hence,
it is essential to keep the time between two questions as
small as possible. For this reason, we introduce an alternative view of our segmentation space based on geodesic
distance transforms (GDT) that recently proved to be extremely efficient in the context of object proposals [15].
We define X = {1, . . . , C}×[0, B]×P(Ω)2 as our state
space and consider the function φGDT : X → S that associates to a vector x = (c, β, Σ+ , Σ− ) ∈ X the segmentation
φGDT (x) ∈ S obtained by computing the geodesic distance
transform on: the cth image channel, blurred with a Gaussian of standard deviation β, with the sets of positive and
negative seeds Σ+ and Σ− respectively. C denotes the number of color channels in the image, and B the maximally
Metropolis-Hastings acceptance probability (Eq. 2) is the
fact that only the ratio of the probabilities p(φGDT (x0 )|Σk )
and p(φGDT (x)|Σk ) appears. Hence, the normalization factor of this probability distribution does not play any role and
we can design this distribution without taking it into consideration. To define this probabilistic term, we propose to
distinguish two cases depending on whether there is at least
one background seed and one foreground seed in Σk .
Figure 2: Segmentation example using geodesic distance
transforms. a) Image from the Berkeley 300 dataset [17].
b) + d) Approximated geodesic distance to outside (orange)
and inside (green) seeds respectively. c) Final segmentation.
allowed standard deviation. The GDT is obtained by computing the shortest distance of every pixel to the set of seed
pixels. Usually, the distance between two neighboring pixels (i.e. the edge weights on the image graph) is defined as a
mixture of the Euclidean distance and the gradient between
these two points. To release the dependence on the seed
placement within the object of interest, we use the squared
intensity difference only. This is particularly important because there is no guarantee that seeds will be in the center of
objects since the user is no longer placing the seeds manually. To generate the final segmentation, each pixel receives
the label of its closest seed. The coefficient β ∈ [0, B] is introduced to control the sensitivity to edges in the image and
implicitly generates an image pyramid during the MCMC
process (see Sec. 2.2.3). Note that any other seed-based interactive segmentation algorithm could be included at this
stage of the framework instead. Our main motivation behind the choice of geodesic distance transforms is the fact
that they can be approximated in linear time [25] and are
hence very fast to compute. The GDT-based segmentation
procedure is further illustrated in Fig. 2.
The Markov chain used for our segmentation sampling
is going to act at the state space level, i.e. on the GDT-based
segmentation parameters and seeds. Eq. 1 becomes
GDT
0
k
k
p(s|Σk ) ∝
(2)
Posterior Probability
The probability p(s|Σk ) states the probability of the segmentation to have been initially picked by the user given
the set of k seeds Σk already revealed by the answers to
the k first questions. An important characteristic of the
(3)
Case 2: Σk+ 6= ∅ and Σk− 6= ∅ Once at least one seed
inside and outside the object is known, we can use the GDT
segmentation algorithm to build a more accurate estimate of
the visual properties of the background and foreground regions. For this, we perform the GDT segmentation based
on the known seeds on each color channel and compute
k
background and foreground intensity histograms H−
and
k
aggregated over the color channels. While building the
H+
histograms, each intensity value is weighted by its inverse
geodesic distance to encode the fact that the confidence decreases with increasing (geodesic) distance to the seeds. To
assign a score to a new segmentation s, we compute similarly the background and foreground histograms of s denoted H− (s) and H+ (s) and measure their mismatch to the
k
k
current estimates H−
and H+
via a chi-squared distance:
p(s|Σk ) ∝
The two next subsections are going to expose our design for
p(φGDT (x)|Σk ) and q(x0 |Σk , x) respectively.
1
,
1 + V ar({I(p), p ∈ s−1 (1)})
where V ar({I(p), p ∈ s−1 (1)}) denotes the variance of
the image values over the set of foreground locations defined by the segmentation s. This variance is summed over
all color channels. Intuitively, we encourage segmentations
which delineate homogeneous regions as foreground.
0
p(φ
(x )|Σ )q(x|Σ , x )
α=
.
p(φGDT (x)|Σk )q(x0 |Σk , x)
2.2.2
Case 1: Σk+ = ∅ or Σk− = ∅ This case, typically occurring during the first questions, corresponds to the absence of
seeds for at least one of the two labels. Note that it occurs
a least for the two first turns, where our knowledge consists
of respectively 0 and 1 seed. We define the segmentation
probability as
2.2.3
1
1+
1
2
P
δ∈{−,+}
χ(Hδk , Hδ (s))
.
(4)
Proposal Distribution
Our segmentations are generated via a set of parameters
x = (c, β, Σ+ , Σ− ) ∈ X sent as input to the GDT segmentation algorithm. The main advantage of this representation
is that the state space X gives a more natural way to move
from a state to another and facilitates the design of the proposal distribution q(.|Σk , x). In practice, we maintain two
sets of seeds Σ+ and Σ− which contain both the already
known and hence reliable seeds included in Σk+ and Σk−
Figure 3: Over-segmentation of the image for the question selection step. The over-segmentation is obtained by
intersecting all ski , 1 ≤ i ≤ N . The most uncertain region
Rj∗ is shown in green and the question pixel pk (shown in
orange) is chosen as the most interior point of Rj∗ .
(fixed seeds) and some other seeds created by the MCMC
process exclusively (mobile seeds). From a given state x, a
state x0 is suggested by drawing uniformly and performing
one of the 5 following moves:
1. Changing image channel: The image channel c is redrawn uniformly.
2. Changing β: The smoothing parameter β is redrawn
uniformly.
3. Adding a seed: A mobile seed is added at a random
(non seed) location and randomly added to Σ+ or Σ− .
4. Removing a seed: A mobile seed is removed from either Σ+ or Σ− (if this does not leave this set empty).
5. Moving a seed: A mobile seed is moved spatially to a
(non seed) pixel according to a normal distribution.
which entirely defines the proposal distribution q(.|Σk , x).
To avoid that the number of seeds diverges, we balance the
probabilities of picking the moves 3. and 4. such that forward and backward moves are equally likely.
2.3. Question Selection
After k questions have been asked and answered (k ≥ 0),
the method described in Sec. 2.2 draws N segmentations
ski , 1 ≤ i ≤ N that approximate the probability distribution p(s|Σk ) that the segmentation s is the answer awaited
by the oracle given the current knowledge (encoded by
Σk ). From these samples, we have to decide on the optimal question to ask to the user. To do so, we first perform an over-segmentation of the image by intersecting all
ski (Fig. 3). This provides a partition of the image into regions (Rj )1≤j≤ρ according to the samples ski . Since the ski
are the only available information, the information carried
by each pixel is constant over an individual segment. The
Figure 4: Representative images from the datasets. From
left to right: (1) Berkeley Segmentation Dataset, (2) Stanford Background Dataset, (3) IBSR Brain Dataset. Note the
diversity in terms of visual content (natural scenes (1,2) vs
medical data (3)) and labels (unsupervised parsing (1) vs
semantic labeling (2,3)).
probability for a region Rj to belong to the object of interest
p(Rj ⊂ ˆs) can be approximated from our samples as
X
1
p(ski |Σk ).
(5)
p(Rj ⊂ ˆs) = PN
k |Σk )
p(s
i
i=1
i|Rj ∈Si
The most uncertain segment is the one for which this probability is the closest to 0.5. However, each segment has
to be weighted by its size to account for the fact that each
pixel will equally benefit from the information carried by
the segment it belongs to. Thus we select the segment Rj∗
maximizing |Rj |(1 − |1 − 2p(Rj ⊂ ˆs)|) as the most informative. At this stage, we could indifferently choose any
location pk ∈ Rj∗ to define the selected question Q(pk ).
In practice, we choose for pk the most interior point of Rj∗
(Fig. 3). Indeed, this results in questions that are in general
easier to answer visually as the most interior point typically
lies far from the edges within the image.
3. Experiments
3.1. Experimental Setup
We describe the parameter settings of our method, that
are the same for the three datasets, in more detail. We use
the CIELab color space and the maximum blur level β, i.e.
the standard deviation of the Gaussian smoothing applied
during the MCMC process (Sec. 2.2.1) is set to B = 6 pixels. The MCMC burn-in step, whose goal is to gain independence from the starting state sk0 , consists of 1000 iterations. 250 Markov chain moves are performed between two
samples drawings to ensure a strong decorrelation between
two consecutive samples. The total number of sampled segmentations at each turn is set to N = 32. It was empirically
found to be sufficient to produce enough variety between
the segmentations such that the over-segmentation does not
become too fine grained without exceeding one second between two questions. The main bottleneck is the computation of the approximate GDT, required at each MCMC
Figure 5: Evolution of the median Dice scores with the
number of questions. The performance is shown on the
three datasets: Berkeley Segmentation Dataset 300 [17],
Stanford Background Dataset [9] and IBSR Brain Dataset.
We also show, on the Stanford background dataset, the individual performance for each class label. The small amount
of mountain labels in the dataset explains the noise on the
corresponding curve.
step, which is linear in the number of pixels. Thus, the total complexity is O(N Kwh). The Markov chain runs on
a downsampled version of the image to speed up computation between the samples, but the drawn segmentations every 250 steps remain computed on the full image to ensure
a fine grained segmentation. The method is evaluated for up
to K = 30 questions. In our experiments, the human user
is simulated by the ground truth segmentation, which allows an extensive evaluation over a large set of images. All
experiments were performed on an Intel i7-4820K 3.7GHz
CPU. The algorithm runs four independent Markov chains
in parallel.
3.2. Results
To measure the quality of a segmentation, we use the
Sørenson index or Dice score [22], which measures the
overlap between the segmentation and the ground truth. We
compare our method against two intuitive baselines. The
first chooses each question by drawing randomly the corresponding pixel location. Our second baseline is an improvement of the first one: as soon as seeds have been
found inside and outside the hidden segmentation, GDTbased segmentation is performed on these seeds. The algorithm uses the resulting pixelwise confidence to select the
question within the area where the label uncertainty is maximal. In other words, this method can be seen as an iterative
refinement of the segmentation border. The experimental
evaluation was performed on three different datasets.
Figure 6: Relationship between size of the hidden segmentation and performance. Shown is a scatter plot of
relative object size (x-axis) vs dice score (y-axis) after 20
questions. Large areas are easier to guess than small ones.
Berkeley Segmentation Dataset 300 The Berkeley Segmentation Dataset 300 [17] is a set of 300 natural images.
For each of them, several ground truth human annotations
defining a parsing of the image into several regions were
collected. This dataset illustrates very well the complexity of finding the area the oracle thinks of, since the parsings are typically different from an annotator to another.
The average number of regions in each individual ground
truth is 20.3. To evaluate our method on this dataset, each
region included in one of the ground truth parsings is selected as oracle segmentation. Figure 5 displays the median
Dice score over all runs when the number of allowed questions increases, for both our method and the aforementioned
baselines. The curves in Fig. 5 start with a plateau at low
performance. This corresponds to the stage where no reliable foreground seed has been found yet, and hence the
sampled segmentations are very noisy since they are only
based on the mobile seeds from the Markov chain. Fig. 5
shows that the median number of questions required to find
a reliable seed is approximately 7. This is interesting since
the dataset contains 20 non-overlapping regions per image
in average. Hence, even if this pool of 20 possible ground
truth segmentations was known in advance, a random seed
placement would still take about 10 questions to find a foreground seed. This demonstrates that our question selection
procedure identifies regions more efficiently.
Stanford Background Dataset The second dataset used
for evaluation is the Stanford Background Dataset [9] which
is a diverse collection of 715 natural images taken from different other datasets. In this case, images have been labeled
according to eight different semantic categories: sky, road,
grass, mountain, water, tree, building and foreground object. This strongly differs from the Berkeley Segmentation
Figure 7: Comparison between our MCMC-based proposals and geodesic object proposals. The heat maps correspond to the frequency of occurrence of each pixel in
the set of sampled segmentations, for the geodesic object
proposal method [15] (middle) and our MCMC sampling
scheme (right). While the former focuses more specifically
on some areas, ours explores more the segmentation space.
Dataset where images were parsed into arbitratry regions
which was intrinsically more ambiguous. The results are
shown on Fig. 5. To give further insights, we also display
the performance for each individual label (Fig. 5) and the relationship between Dice score after 20 questions and size of
the target segmentation (Fig. 6). Bigger segmentations (respectively segmentations of a typically predominant class)
appear to be easier to guess than smaller ones. This correlation between object size and segmentation quality is rather
natural and directly comes from the type of questions that
are asked. This point is discussed in more details in Sec. 3.3.
IBSR MR Brain Data To assess further the generality of
our method, we evaluated it on the medical IBSR MR brain
datasets composed of 18 magnetic resonance brain scans
together with labels of 4 different brain structures. We decomposed the scans into 2D slices to have a larger number of runs for our evaluation. The data and their manual
segmentations were provided by the Center for Morphometric Analysis at Massachusetts General Hospital and are
available at http://www.cma.mgh.harvard.edu/
ibsr/. Labeling the brain structures is a very complex task
which requires a high expertise. The results of our method
on the IBSR dataset can be seen in Fig. 5. As expected, the
absolute performance is lower than on the two other datasets
which had a richer visual content, but a significant improvement after 20 questions remains achieved.
Comparison with geodesic object proposals Finally, we
compared our MCMC-based sampling of segmentations described in Sec. 2.2 with the recent geodesic object proposals
(GOP) introduced by Kr¨ahenb¨uhl and Koltun [15], whose
goal is to suggest objects of interest in an image. By using the available (pre-trained) code for GOP, we sample for
each image both N object proposals with GOP and N segmentations with our sampling method, where N is automatically inferred by the GOP algorithm. We report the Jaccard
index of the sample matching the ground truth best (ABO
score [15]). The GOP obtain an average score of 66.0 on the
Stanford Background Dataset and 43.1 on the ISBR, while
Figure 8: Qualitative results. Examples of oracle segmentation and the corresponding segmentation guess after
20 questions. From left to right: Berkeley Segmentation
Dataset, Stanford Background Dataset, IBSR Brain Dataset.
our method obtains respectively 59.6 and 51.0. GOP appear
to have an advantage on the Stanford Background Dataset,
presumably because they rely on a seed placement technique which was learned precisely on outdoor scenes. On
the contrary, our method does not rely on any offline training step, i.e. makes no assumption on the image content,
and still performs better on brain images which illustrates
the flexibility of our framework. Figure 7 shows another interesting difference between the two proposal frameworks.
Without prior knowledge on the current image, the segmentations proposed by GOP are more redundant than ours,
which is a good property to propose the likeliest object but
suits less the context of an interactive guessing scenario.
Dataset
Berkeley
Stanford
IBSR
Method
Random
Uncertainty
Ours
Random
Uncertainty
Ours
Random
Uncertainty
Ours
10Q
30.9 / 19.0
32.5 / 17.5
34.7 / 23.8
49.8 / 56.5
50.6 / 57.6
52.6 / 63.9
43.1 / 47.8
41.7 / 45.8
51.5 / 53.7
20Q
41.7 / 43.7
46.9 / 55.0
48.8 / 62.0
60.2 / 69.8
61.9 / 71.4
63.9 / 75.8
52.0 / 57.6
46.2 / 50.7
58.4 / 60.1
30Q
48.9 / 56.2
55.9 / 70.3
56.1 / 73.2
65.6 / 74.7
67.9 / 77.7
67.9 / 79.8
58.1 / 63.6
49.2 / 53.3
62.9 / 64.7
Table 1: Quantitative Results. Mean and median Dice
score after 10, 20 and 30 questions on the three datasets
for our method and the two baselines described in Sec. 3.2
Inaccurate Answers To explore the robustness of the
proposed method with respect to inaccurate answers to the
questions, we tried an alternative experimental setting on
the SBD dataset such that, when a question is posed within
3 pixels of the border of the desired object, the answer is
randomly decided (50% yes 50% no). This happens about
20% of the time, mainly on the later questions. In average, 1.2 additional questions are needed to achieve the same
Dice score as with perfect answers. Hence, the proposed
method is reasonably robust to wrong answers.
Figure 9: Evolution of the segmentation belief with the number of answers collected. a) shows the original image and
the segmentation chosen by the oracle. Below, f) shows the final segmentation proposed by the computer after 30 questions.
The images on the right show the evolution of the segmentation belief with the number of seeds collected (b: 0, c: 10, d: 20,
e: 30). Both the intersection of all sampled segmentations (i.e. the over-segmentation described in 2.3 and shown in Fig. 3)
and a heat map of the pixel likelihood are shown, respectively on the top and bottom row. The heat maps are generated by
computing, for each pixel, the proportion of sampled segmentations that contained it (similarly to Eq. 5). The positive (resp.
negative) seeds collected from the answers are shown in green (resp. orange). Note how the questions are progressively asked
towards the left where the uncertainty is the highest, so that the whole series of buildings can be eventually segmented.
3.3. Discussion
We start this discussion with an example to illustrate further the behavior of the method. Figure 9 shows a typical instance of the scenario. The oracle thinks of an object, which
is in this case the row of buildings in the background. It
is a challenging target object for a segmentation algorithm
due to the complexity of its structure and the fact that it
corresponds to a rather high-level semantic. The computer
generates a first set of possible segmentations. Since no
seeds are available yet, uniform regions like sky or pavement are favored at this stage. The first question is asked
at the location bringing the most information according to
the drawn samples. As answers from the oracle are progressively collected, we can see that the space of sampled
segmentations changes drastically. The algorithm suggests
questions at boundary locations, and expands step by step
the initial guess towards the left. A good estimate of the oracle segmentation is eventually provided after 30 questions.
The results demonstrate overall that our method has
more difficulties to segment small objects, and in particular to find a first foreground seed in these cases. While it
seems relatively intuitive that it takes longer to find small
areas than big ones, the nature of the questions we ask is
partially responsible for this as well. We investigated this by
changing the kind of question asked. For instance, a more
efficient solution to find small objects would be to show a
highlighted area to the user and ask “Is the target segmentation fully inside the shown area?”. If the answer is yes,
all unhighlighted pixels could be added to the outside seeds
and the search space would be greatly reduced. However, a
no would provide almost no information since it would only
state that at least one (unknown) pixel of the object is outside the shown area. Hence, this kind of questions suffers
from the opposite problem: they are efficient to find small
objects but poor at finding large ones, and have the additional drawback that a negative answer is practically very
difficult to leverage. This is mainly what motivated us to
choose the location-based question type, but we believe that
a combination of both types of questions would be ideally
suited to find both small and large objects in the image.
4. Conclusion
We introduced a method able to guess a segmentation in
an arbitrary image by asking simple binary questions to the
user. The questions are computed by approximating a posterior probability distribution over the set of segmentations
with a MCMC sampling algorithm building on geodesic
distance transform segmentation. Our method shows to be
tractable for practical use with a waiting time between two
questions of less than one second. No assumption about the
type of the given image is necessary as the framework does
not rely on any offline training step. The experimental evaluation performed on three very different datasets demonstrates that the approach provides an overall efficient solution to this problem. Our future directions of work include
investigating the feasibility of mixing several question types
and automatically learning the most relevant ones for the
hidden segmentation. Since answering binary questions is
plausibly more attractive than manual annotation, we are
also considering using this method to generate a lot of training data for seed placements in a crowdsourcing fashion.
Acknowledgements With the support of the Technische
Universit¨at M¨unchen - Institute for Advanced Study, funded
by the German Excellence Initiative and the European
Union Seventh Framework Programme under grant agreement no 291763, and with the partial support of the DFGfunded Collaborative Research Centre 824: Imaging for Selection, Monitoring and Individualization of Cancer Therapies.
References
[1] 20q.net: The neural-net on the internet (www.20q.net).
[2] Akinator the web genie (http://en.akinator.com).
[3] D. Batra, A. Kowdle, D. Parikh, J. Luo, and T. Chen. icoseg:
Interactive co-segmentation with intelligent scribble guidance. In CVPR 2010, pages 3169–3176. IEEE, 2010.
[4] Y. Boykov and M.-P. Jolly. Interactive graph cuts for optimal
boundary amp; region segmentation of objects in n-d images.
In ICCV 2001, volume 1, pages 105–112 vol.1, 2001.
[5] S. Branson, C. Wah, F. Schroff, B. Babenko, P. Welinder,
P. Perona, and S. Belongie. Visual recognition with humans
in the loop. In K. Daniilidis, P. Maragos, and N. Paragios, editors, Computer Vision ECCV 2010, volume 6314 of Lecture
Notes in Computer Science, pages 438–451. Springer Berlin
Heidelberg, 2010.
[6] J. Carreira and C. Sminchisescu. CPMC: Automatic Object Segmentation Using Constrained Parametric Min-Cuts.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012.
[7] I. Endres and D. Hoiem. Category-independent object proposals with diverse ranking. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 36(2):222–234,
Feb 2014.
[8] A. Fathi, M. F. Balcan, X. Ren, and J. M. Rehg. Combining self training and active learning for video segmentation.
In Proceedings of the British Machine Vision Conference
(BMVC 2011), volume 29, pages 78–1, 2011.
[9] S. Gould, R. Fulton, and D. Koller. Decomposing a scene
into geometric and semantically consistent regions. In Computer Vision, 2009 IEEE 12th International Conference on,
pages 1–8. IEEE, 2009.
[10] L. Grady. Random walks for image segmentation. Pattern
Analysis and Machine Intelligence, IEEE Transactions on,
28(11):1768–1783, Nov 2006.
[11] L. Grady, M.-P. Jolly, and A. Seitz. Segmentation from a
box. In Computer Vision (ICCV), 2011 IEEE International
Conference on, pages 367–374. IEEE, 2011.
[12] C. Gu, J. Lim, P. Arbelaez, and J. Malik. Recognition using
regions. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on, pages 1030–1037, June
2009.
[13] A. Kowdle, D. Batra, W.-C. Chen, and T. Chen. imodel: interactive co-segmentation for object of interest 3d modeling.
In Trends and Topics in Computer Vision, pages 211–224.
Springer, 2012.
[14] A. Kowdle, Y.-J. Chang, A. Gallagher, and T. Chen. Active
learning for piecewise planar 3d reconstruction. In CVPR,
2011, pages 929–936. IEEE, 2011.
[15] P. Kr¨ahenb¨uhl and V. Koltun. Geodesic object proposals.
In D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, editors,
Computer Vision ECCV 2014, volume 8693 of Lecture Notes
in Computer Science, pages 725–739. Springer International
Publishing, 2014.
[16] V. Lempitsky, P. Kohli, C. Rother, and T. Sharp. Image
segmentation with a bounding box prior. In ICCV, number
MSR-TR-2009-85, 2009.
[17] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database
of human segmented natural images and its application to
evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int’l Conf. Computer Vision, volume 2, pages 416–423, July 2001.
[18] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H.
Teller, and E. Teller. Equation of state calculations by fast
computing machines. The Journal of Chemical Physics,
21(6):1087–1092, 1953.
[19] C. Nastar, M. Mitschke, and C. Meilhac. Efficient query
refinement for image retrieval. In CVPR 1998, pages 547–
552, Jun 1998.
[20] B. Price, B. Morse, and S. Cohen. Geodesic graph cut for
interactive image segmentation. In CVPR 2010, pages 3161–
3168, June 2010.
[21] M. Sadeghi, G. Tien, G. Hamarneh, and M. S. Atkins.
Hands-free interactive image segmentation using eyegaze.
In SPIE Medical Imaging, pages 72601H–72601H. International Society for Optics and Photonics, 2009.
[22] T. Sørensen. A method of establishing groups of equal amplitude in plant sociology based on similarity of species and
its application to analyses of the vegetation on danish commons. Biol. skr., 5:1–34, 1948.
[23] C. Straehle, U. Koethe, G. Knott, K. Briggman, W. Denk,
and F. Hamprecht. Seeded watershed cut uncertainty estimators for guided interactive segmentation. In CVPR, 2012,
pages 765–772, June 2012.
[24] K. Tieu and P. Viola. Boosting image retrieval. In CVPR,
2000. Proceedings., volume 1, pages 228–235 vol.1, 2000.
[25] P. J. Toivanen. New geodosic distance transforms for grayscale images. Pattern Recognition Letters, 17(5):437–450,
1996.
[26] S. Tong and E. Chang. Support vector machine active learning for image retrieval. In ACM International Conference
on Multimedia, pages 107–118, New York, NY, USA, 2001.
ACM.
[27] Z. Tu and S.-C. Zhu. Image segmentation by data-driven
markov chain monte carlo. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 24(5):657–673, May
2002.
[28] C. Wah, S. Branson, P. Perona, and S. Belongie. Multiclass
recognition and part localization with humans in the loop. In
Computer Vision (ICCV), 2011 IEEE International Conference on, pages 2524–2531, Nov 2011.